Fusing representations for code-mixed low resource language processing

Problem?

• Low resource language data being generated at a large scale now.
   
• Major work has been done only for English-x language code-mixed language models.
   
• Untapped potential of low resource to low resource code-mixed language data.

Notable Work

• Dravidian-CodeMix - FIRE 2021
   

• GCM: A Toolkit for Generating Synthetic Code-mixed Text (Rizvi et al. 2021)
   

• Transfer learning based code-mixed part-of-speech tagging using character level representations for Indian languages
  (Madasamy et al. 2021)

Pre-processing for the z-vector

• Text cleaning using Ekphrasis
   
• POS tagging to create a "(subject, object, verb, modifier)" pointer
   
•      is passed through a LSTM+Word Attention model to create the input to GAN:

Impact?

• Low resource code-mixed language processing
• Transfer Learning to English-x code-mixed tasks
• Social computing based Computational Social Science Tasks
• Future tasks may include probing these models for possible learned biases

Thank you!